textmachine/backend/internal/pipeline/flagseverity_test.go
2026-09-15 14:18:58 +03:00

286 lines
14 KiB
Go

package pipeline
import (
"fmt"
"go/ast"
"go/parser"
"go/token"
"go/types"
"os"
"strconv"
"strings"
"testing"
"textmachine/backend/internal/llm"
)
// flagseverity_test.go pins the chapter passport's "worst flag" ranking.
//
// ⚠ WHY THIS FILE EXISTS AT ALL. flagReasonSeverity had no test of any kind, and the shape of the bug it
// carried is the one this project keeps meeting: a hand-written list that stopped tracking the constants it
// lists. `off_target_lang` — the flag an entire pack was built to raise — was never added to the switch, so
// it fell to the default and became the MOST BENIGN reason in the passport, below `length`. A chapter that
// came back in the wrong language reported «length» as its worst problem. Nothing was wrong with the code;
// the list was simply not tied to anything.
// TestEveryFlagReasonIsRanked reads the FlagReason constants OUT OF THE SOURCE rather than retyping them
// here, because a hand-written list in the test would rot exactly like the hand-written list in the code —
// and would then certify the rot. This is the guarantee that the next flag someone adds cannot land in the
// unknown bucket silently: it fails here on the day it is declared, not on the day it is misreported.
func TestEveryFlagReasonIsRanked(t *testing.T) {
// ⚠ EVERY FILE OF THE PACKAGE, not disposition.go alone — the first version of this test named one
// file, and a planting that declared a FlagReason anywhere else walked straight past it while the test
// still reported «every declared reason is ranked». The guarantee was one file narrower than its own
// sentence, which is the defect class this file exists to prevent, committed by the file itself.
entries, err := os.ReadDir(".")
if err != nil {
t.Fatalf("read the package directory: %v", err)
}
var sources []string
for _, e := range entries {
n := e.Name()
if e.IsDir() || !strings.HasSuffix(n, ".go") || strings.HasSuffix(n, "_test.go") {
continue
}
sources = append(sources, n)
}
if len(sources) < 10 {
t.Fatalf("only %d source files found — the walk stopped seeing the package: %v", len(sources), sources)
}
fset := token.NewFileSet()
var decls []ast.Decl
for _, name := range sources {
f, perr := parser.ParseFile(fset, name, nil, 0)
if perr != nil {
t.Fatalf("parse %s: %v", name, perr)
}
decls = append(decls, f.Decls...)
}
var found int
for _, d := range decls {
gd, ok := d.(*ast.GenDecl)
if !ok || gd.Tok != token.CONST {
continue
}
for _, sp := range gd.Specs {
vs, ok := sp.(*ast.ValueSpec)
if !ok {
continue
}
// ⚠ THE TYPE MUST BE WRITTEN ON THIS SPEC. An earlier version carried the previous spec's type
// forward, which is not how Go types a const block: a spec that has a VALUE and no type is its
// own UNTYPED constant and inherits nothing. That version raised a FALSE ALARM — an ordinary
// string constant placed after a FlagReason spec in the same block was accused of being an
// unranked flag, and a doc-URL constant would have failed this test with a message about the
// chapter passport. A test that accuses the innocent is not enforcing its own sentence, and this
// one exists to say what IS a declared reason.
//
// ⚠ AND BOTH GO SPELLINGS COUNT. `X FlagReason = "…"` writes the type in the spec; the equally
// ordinary `X = FlagReason("…")` writes it as a CONVERSION, leaving vs.Type nil — and a walk
// that looked only at vs.Type was blind to it, which an adversarial pass exhibited as a pair:
// the typed spelling of one declaration RED, the conversion spelling of the same declaration
// SURVIVED. The sentence that used to stand here — «a missed declaration would trip the found
// floor rather than pass silently» — was simply false: one missing constant leaves the count at
// 16 against a floor of 15, and nothing trips.
if len(vs.Values) == 0 {
continue
}
// ⚠ AND A THIRD SPELLING, which this walk used to drop on the floor: the value can be a
// conversion of a constant from ANOTHER package — `X FlagReason = FlagReason(llm.Y)` — used so
// that a vocabulary the wire, the checkpoint, the flag and the operator all read has ONE
// carrier. There is no literal to unquote there, the old code hit `lit == nil` and `continue`d,
// and both flags declared that way landed in the unknown bucket while THIS TEST STAYED GREEN.
// Measured when it happened: 18 ranks against 16 constants the walk could see.
//
// So an unreadable value is now a FAILURE, not a skip. A silent `continue` in an
// exhaustiveness test is the same defect the test exists to catch, one level up.
decl, isDecl := flagReasonValue(vs)
if !isDecl {
continue
}
val, err := decl.resolve()
if err != nil {
t.Errorf("%s: %v", vs.Names[0].Name, err)
continue
}
if val == "" {
continue // reasonOK — the sentinel for "not a flag", deliberately unranked
}
found++
if _, ranked := flagSeverity[FlagReason(val)]; !ranked {
t.Errorf("%s (%q) has no entry in flagSeverity: it would rank %d — BELOW every real diagnosis, "+
"so it can never be reported as a chapter's worst flag", vs.Names[0].Name, val, severityUnknown)
}
}
}
// The reader of this test must know it actually read something: a parse that silently matched nothing
// would pass forever and guard nothing — the same failure class one layer up.
if found < 15 {
t.Fatalf("only %d FlagReason constants were read out of the package's %d source files — the walk "+
"stopped matching the source, and this test is no longer checking anything", found, len(sources))
}
if len(flagSeverity) != found {
t.Errorf("flagSeverity has %d entries for %d declared reasons — a rank exists for a reason that no "+
"longer does", len(flagSeverity), found)
}
}
// TestOffTargetLangIsRankedWithTheContentFailures is the specific harm, asserted as a number and not as a
// membership: the flag must out-rank the budget symptoms, because those are what it was losing to.
func TestOffTargetLangIsRankedWithTheContentFailures(t *testing.T) {
off := flagReasonSeverity(string(FlagOffTargetLang))
if off == severityUnknown {
t.Fatal("off_target_lang fell to the unknown bucket — the exact defect this file was written for")
}
for _, milder := range []FlagReason{FlagLength, FlagEmpty, FlagGlossaryMiss, FlagSanitizerStripped} {
if off >= flagReasonSeverity(string(milder)) {
t.Errorf("off_target_lang (%d) must be WORSE than %s (%d): a chapter in the wrong language is not "+
"a budget symptom", off, milder, flagReasonSeverity(string(milder)))
}
}
// It is a content failure, ranked with the echo it sits beside in the classifier.
if off != flagReasonSeverity(string(FlagCJKArtifact)) {
t.Errorf("off_target_lang (%d) and cjk_artifact (%d) are the same class — «the output is not a "+
"translation of this text into the asked-for language» — and must rank together",
off, flagReasonSeverity(string(FlagCJKArtifact)))
}
// A refusal still outranks both: the provider said no, which is a different and worse fact.
if off <= flagReasonSeverity(string(FlagHardRefusal)) {
t.Error("a hard refusal must still outrank an off-target completion")
}
}
// isFlagReasonIdent reports whether an expression names the FlagReason type — as a spec's type or as the
// function of a conversion. One helper for both spellings, so the walk cannot learn one and forget the other.
func isFlagReasonIdent(e ast.Expr) bool {
id, ok := e.(*ast.Ident)
return ok && id.Name == "FlagReason"
}
// flagValueDecl is a FlagReason constant declaration the walk has recognised, in whichever of the three
// spellings it was written.
type flagValueDecl struct {
lit *ast.BasicLit // `X FlagReason = "y"` or `X = FlagReason("y")`
conv string // `X FlagReason = FlagReason(pkg.Y)` — the rendered source expression
}
// convertedFlagValues resolves the conversion spelling. The keys are SOURCE EXPRESSIONS and the values
// are compile-time references to the very constants those expressions name, so the map cannot drift in
// VALUE — only in membership, and a member it lacks fails loud below instead of being skipped.
var convertedFlagValues = map[string]FlagReason{
"llm.CutBySelfDeadline": FlagReason(llm.CutBySelfDeadline),
"llm.CutByParent": FlagReason(llm.CutByParent),
"llm.CutByConnection": FlagReason(llm.CutByConnection),
}
func (d flagValueDecl) resolve() (string, error) {
if d.lit != nil {
if d.lit.Kind != token.STRING {
return "", fmt.Errorf("declared with a non-string value %s — a FlagReason is a string", d.lit.Value)
}
return strconv.Unquote(d.lit.Value)
}
v, ok := convertedFlagValues[d.conv]
if !ok {
return "", fmt.Errorf("declared as a conversion of %s, whose value this test cannot read from the "+
"source; add it to convertedFlagValues (a compile-time reference, not a retyped string) so the "+
"exhaustiveness check can see it instead of skipping it", d.conv)
}
return string(v), nil
}
// flagReasonValue recognises a FlagReason declaration in any of its three spellings and reports how its
// value can be read. It returns false for a spec that is not a FlagReason at all — an ordinary string
// constant sitting in the same block is not this test's business.
func flagReasonValue(vs *ast.ValueSpec) (flagValueDecl, bool) {
typed := isFlagReasonIdent(vs.Type)
if lit, ok := vs.Values[0].(*ast.BasicLit); ok {
return flagValueDecl{lit: lit}, typed
}
call, isCall := vs.Values[0].(*ast.CallExpr)
if !isCall || !isFlagReasonIdent(call.Fun) || len(call.Args) != 1 {
return flagValueDecl{}, false
}
if lit, ok := call.Args[0].(*ast.BasicLit); ok {
return flagValueDecl{lit: lit}, true
}
return flagValueDecl{conv: types.ExprString(call.Args[0])}, true
}
// TestTheSeverityTableMeansWhatItsCommentsSay pins the ORDER, not the membership. The exhaustiveness
// test beside it proves every reason has a rank; it says nothing about what the ranks are, so the whole
// table could be re-ordered — `cancelled` declared a chapter's worst problem, a lost chunk declared its
// mildest — and the battery would stay green.
//
// Every assertion below is a sentence the table already writes about itself, turned into a check. That
// is the point: a rank is a claim, and a claim with no carrier is a comment.
func TestTheSeverityTableMeansWhatItsCommentsSay(t *testing.T) {
rank := func(r FlagReason) int {
s, ok := flagSeverity[r]
if !ok {
t.Fatalf("premise broken: %s carries no rank, so the comparisons below compare nothing", r)
}
return s
}
// The control first: the table is not empty and not all one value, or every «is milder than» below
// would hold trivially.
seen := map[int]bool{}
for _, s := range flagSeverity {
seen[s] = true
}
if len(flagSeverity) < 5 || len(seen) < 3 {
t.Fatalf("premise broken: %d reasons across %d distinct ranks — an ordering test needs an order",
len(flagSeverity), len(seen))
}
// «The mildest mark there is» — a passport that reported `cancelled` as a chapter's worst problem
// would hide a durable finding behind a state the next run erases.
for r, s := range flagSeverity {
if r == FlagCancelled {
continue
}
if rank(FlagCancelled) <= s {
t.Fatalf("`cancelled` must be milder than every other mark — it is the only one the engine "+
"removes by itself — but it ranks %d against %s at %d", rank(FlagCancelled), r, s)
}
}
// «They rank together because they are the same thing to a reader — the chunk is lost and the money
// is spent.» A lost connection is not a member: it reaches no disposition, so it wears no rank.
//
// ⚠ THE THIRD MEMBER JOINED WITH A CLAIM OF ITS OWN (row 291): `retry_unaffordable` is a position that
// paid for an attempt and whose re-attack a ceiling refused, so to a reader it is the same sentence —
// the chunk is lost and the money is spent — and the table says so. Asserted here because a rank
// written in a comment and nowhere else is a claim with no carrier, and a third member added to the
// comment alone would have left the trio's equality measured on two of three.
for _, r := range []FlagReason{FlagAttemptTimeout, FlagRetryUnaffordable} {
if rank(FlagDecodeError) != rank(r) {
t.Fatalf("the paid-and-nothing-came-back reasons must share one rank: decode=%d %s=%d",
rank(FlagDecodeError), r, rank(r))
}
}
// «It IS one of those two, plus the fact that the remedy could not be bought» — a chapter whose unit
// died for lack of money must not report `length` as its worst problem and send a person to fix a
// budget formula instead of topping up.
if rank(FlagRetryUnaffordable) >= rank(FlagLength) || rank(FlagRetryUnaffordable) >= rank(FlagEmpty) {
t.Fatalf("a re-attack nobody could afford must out-rank the budget symptom it supersedes: "+
"unaffordable=%d length=%d empty=%d", rank(FlagRetryUnaffordable), rank(FlagLength), rank(FlagEmpty))
}
// «An unrecognised string must not out-rank a diagnosis the engine actually made.»
for r, s := range flagSeverity {
if severityUnknown <= s {
t.Fatalf("severityUnknown (%d) must be milder than every diagnosis the engine makes, but %s "+
"ranks %d", severityUnknown, r, s)
}
}
// «Ranked below a budget symptom (the chunk is not lost)» — a stripped chunk SHIPPED.
if rank(FlagSanitizerStripped) <= rank(FlagLength) || rank(FlagSanitizerStripped) <= rank(FlagEmpty) {
t.Fatalf("an auto-cleaned chunk that SHIPPED must be milder than a budget symptom that lost one: "+
"stripped=%d length=%d empty=%d", rank(FlagSanitizerStripped), rank(FlagLength), rank(FlagEmpty))
}
// «Ranked with the deterministic content failures, above a mere budget symptom» — a DROPPED
// contaminated output is unreadable as shipped.
if rank(FlagSanitizerDefect) >= rank(FlagLength) {
t.Fatalf("a contaminated output that was DROPPED must be more severe than a budget symptom: "+
"defect=%d length=%d", rank(FlagSanitizerDefect), rank(FlagLength))
}
}